[python] Generated output differs between identical runs without --embed-auditing-info

Open
#412 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
45/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
python
Domain
build-system

Research direction

Start by reading the three causes in the issue and inspect src/nunavut/lang/py/templates/Namespace.j2, src/nunavut/lang/py/__init__.py around filter_pickle, and src/nunavut/_namespace.py around get_nested_types(). Check the Python generation tests for an existing output comparison pattern. Done means repeated runs without --embed-auditing-info produce identical output, with regression coverage for the timestamp, gzip metadata, and namespace ordering.

Written by the indexing model from the issue text.

Description

Python

Summary

Running the same nnvg --target-language=py command twice produces different files, even without --embed-auditing-info. Generated output is supposed to be identical when that flag isn't given (#324 reversed the flag so that reproducible output is the default). C output is identical between runs; Python output is not.

Reproduce

On main at ed853f9, in ghcr.io/opencyphal/toxic:tx26.4.1:

PRDT=submodules/public_regulated_data_types/uavcan
nnvg $PRDT --target-language=py --outdir a
nnvg $PRDT --target-language=py --outdir b
diff -rq a b | wc -l     # 244 of 245 files differ (only nunavut_support.py matches)

The same two runs with --target-language=c produce 0 differing files out of 176. C++ was not checked.

Causes

There are three independent causes. All three must be fixed for the output to be reproducible.

  1. The namespace file's timestamp ignores --embed-auditing-info (affects all 69 __init__.py). src/nunavut/lang/py/templates/Namespace.j2:8 writes # Generated at: {{ now_utc }} UTC unconditionally. Python's type modules (py/templates/base.j2) and the C and C++ base.j2 templates write it only inside {%- if nunavut.embed_auditing_info %}.
  2. The embedded model blob records the time (affects all 175 type modules). filter_pickle (src/nunavut/lang/py/__init__.py:355) calls gzip.compress(pickle.dumps(x, protocol=4)). Without an mtime argument, gzip writes the current time into its header, so each module's _MODEL_ constant changes on every run. Decoding the blobs from both runs gives identical pickled payloads. Only the 4-byte gzip MTIME header field differs (for example 1792505892 against 1792505915). Nothing else in the type modules differs.
  3. Namespace file import order is nondeterministic (affects 30 of 68 uavcan __init__.py). Ignoring the timestamp line, 30 namespace files still differ, always in the order of their from … import … lines and never in which types they list. Namespace.get_nested_types() (src/nunavut/_namespace.py:904) returns types in insertion order, and insertion order follows the order the DSDL files were read (set iteration plus parallel parse-job completion).

Expected

Without --embed-auditing-info, running the same command twice produces byte-for-byte identical Python output, as it does for C. With the flag, volatile metadata such as the timestamp may appear, as it does in the other languages.

Possible fixes

  1. Wrap the Generated at line in py/templates/Namespace.j2 in {%- if nunavut.embed_auditing_info %}, matching py/templates/base.j2.
  2. Pass mtime=0 to gzip.compress in filter_pickle. Consumers (_restore_constant_) only call gzip.decompress, which ignores the field.
  3. Emit namespace members in a stable order, e.g. sorted by full name and version, in get_nested_types() or in the template.

A regression test could generate the same input twice into two directories, without --embed-auditing-info, and assert that the trees are identical, for every target language.

Found while working on #403. It doesn't block #403, which needs semantic idempotence, not byte-for-byte reproducibility.

Dominant language
Python
Stars
48
Forks
26
Avg merge
1d 18h
Merged PRs (30d)
9

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from OpenCyphal/nunavut

All issues in OpenCyphal/nunavut

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.